Agents as 1960s Office RPG

I’ve been thinking a lot lately, as have many others, about where humans fit in the Internet services value chain given the recent explosion in LLM-based coding assistant tools. To answer that one must have a sense of how useful (or not) these tools are. And to answer that requires one to go on a journey of discovery which inevitably leads to the question: “How do I effectively manage context windows?”

One possible answer: treat programming with AI coding assistants like a 1960s workplace role playing game.

The gist is to treat tools like Claude Code as pieces in an RPG, and embrace 1960s organizational dynamics:

• rigid roles & responsibilities

• rigid org chart

• interdepartmental memos

• complete(-ish) specs before the first line of code is written

Independent instances of Claude Code or your tool of choice serve as the various departments and team members[0]. Instead of having “Claude the code writer” you have “Claude the...

• system architect

• API designer

• technical documentation writer

• implementation gap analyst

• security engineer

• QA engineer

• infrastructure engineer

• software engineer

• reliability engineer

• etc.

each with its own prompts and context windows.

A directory of Markdown files becomes your interoffice mail system, your message bus. Instead of prompting Claude to “write code that does x” you prompt the Claude instance with the “task planner” persona to generate checklists in Markdown files of all the individual software components which will need to be written in order to produce code necessary to comply with the spec and documentation, and prompt a separate Claude instance with the “software engineer” persona to iterate through those checklists. Instead of trying to wrangle a wayward Claude instance that’s writing code you don’t want, you first prompt the “gap analyst” persona to output a list of spec and documentation gaps which can be filled prior to generating code so those instances are more likely to stay on track.

I don’t think this idea of distinct personas is unique, and indeed things like Claude Code’s /agent suggest others have already implemented these ideas. But “agentic coding” currently has a branding problem, because newcomers don’t know what it means. I’ve certainly enjoyed my internal dialectic reasoning journey as I’ve tried to understand the changing roles of tools and humans in software production, but some people just want to jump in and get things done faster.

So I just think of this as the elevator pitch for LLM-based coding assistants and agentic coding tools: Simulate a 1960s office. In a loop.

[0] -- This idea of the 1960s office RPG predates the release of agent support in Claude Code--for me it was just a dozen instances of claude in separate tmux windows--but its release, and conversations I’ve had with others, showed me that many of us are taking own own paths and ending up at similar conclusions.